/compact fails with UnexpectedModelBehavior on GitHub Copilot models (non-streaming summarization call)

Description

/compact with the summarization strategy always fails when the summarization model is a GitHub Copilot model. The compaction reports success but achieves 0% reduction, so context is never actually reclaimed.

The model produces a perfectly good summary — but it's discarded during response validation and never reaches the history.

Steps to reproduce

1. Use a Copilot model
2. Leave summarization_model empty so it falls back to the global model.
3. Fill up context, then run /compact

Actual behaviour

Compacting 132 messages using summarization strategy... (~56,502 tokens) Summarizing 18 older messages Compaction failed: [SummarizationError] LLM call failed during summarization: [UnexpectedModelBehavior] Invalid response from openai chat completions endpoint: 2 validation errors for ChatCompletion choices.0.index Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType] object Input should be 'chat.completion' [type=literal_error, input_value=None, input_type=NoneType] Done! History: 132 → 132 messages via summarization Tokens: 56,502 → 56,502 (0.0% reduction)

Root cause

Three things line up:

1. Copilot returns a non-spec response envelope. Captured off the wire from https://api.enterprise.githubcopilot.com/chat/completions:

top-level keys : ['choices', 'copilot_usage', 'created', 'id', 'model', 'usage']
choice[0] keys : ['finish_reason', 'message']

Missing the top-level "object": "chat.completion" discriminator and the per-choice "index". The message content itself is complete and valid.

2. The OpenAI SDK doesn't validate. It builds ChatCompletion via model_construct (openai/_models.py:1001), so both absent fields silently become None.

3. pydantic-ai validates strictly, but only on the non-streaming path. OpenAIChatModel._process_response calls _validate_completion (pydantic_ai/models/openai.py:1203), which runs a strict _ChatCompletion.model_validate(...) and raises UnexpectedModelBehavior at line 1205. None fails both int and the 'chat.completion' literal.

Start with the /compact summarization path and inspect OpenAIChatModel._process_response and _validate_completion in pydantic_ai/models/openai.py, alongside openai/_models.py:1001. Reproduce the non-streaming GitHub Copilot response and verify that compaction accepts its valid message, updates history, and reports a reduction instead of 0%.